Legal

AI Disclosure

Effective 10 July 2026 · Version 1.0

Whizz Pulse uses speech recognition and large language models to analyze recorded customer-service calls. The people on those calls — customers and agents — deserve a plain statement of what the AI does, what it is prevented from doing, and what can be asked of the companies that use it. This page is that statement, written for call participants, customers, and regulators alike.

1. What the AI does, stage by stage

• Transcribe. A best-in-class speech-to-text model, run by our speech-recognition subprocessor, converts the uploaded recording into text with word timestamps and diarization — separating who spoke when within that one call. Diarization is acoustic separation inside a single recording; it does not identify who a voice belongs to, and no voiceprint is stored or matched across calls.
• Redact. Before any language model reads the transcript, deterministic code — not AI — scrubs email addresses, payment-card numbers, and long digit strings, replacing them with placeholders. Every later stage works on the redacted text.
• Analyze. A frontier language model, running on our cloud infrastructure, reads the redacted transcript and produces a summary, conversational sentiment overall and per speaker, topics, pain points and service gaps, predicted CSAT, escalation and churn risk, compliance flags, and coaching suggestions. The analysis is text-based: it is derived from the words spoken, not from voice tone, biometrics, or physiological signals.
• Score.The same model scores the call against the customer's QA scorecard — criteria and weights the customer defines. Every criterion score must be justified with verbatim quotes from the transcript, and each carries a confidence value; low-confidence scores flag the whole result for human review. The overall score is a weighted average computed arithmetically from the criterion scores — no model involvement.

2. Evidence, not verdicts

A score no one can inspect is a score no one can contest. Every QA result carries per-criterion scores, the weights used, the model's reasoning, and the exact transcript quotes behind each score — in the call's original language. This is the record a supervisor uses to review an outcome, and the record an agent can be shown when a score is discussed.

3. Human review

Whizz Pulse scores and recommends; it does not act. It cannot discipline an agent, contact a customer, or change anything outside its own dashboard. Scores where the model reports low confidence are explicitly flagged needs review. Our terms require customers not to take adverse employment action against an agent based solely on an automated score without human review.

4. Who is affected, and what they can ask for

• Agents whose calls are scored can ask their employer for the evidence behind any score — it exists for every scored call — and for human review of automated results.
• Callers whose conversations are analyzed have data rights (access, deletion, objection) exercised through the company that recorded the call, per our Privacy Policy. Requests sent to support@whizztech.ai are routed to the responsible organization.

5. Regulatory posture, in plain language

• Call-recording law — consent and notice requirements for recording calls vary by jurisdiction and sit with the company doing the recording. Whizz Pulse only ever receives recordings after the fact; our terms require customers to warrant the recordings were made lawfully.
• UAE PDPL and GDPR — we act as processor for call content, under a data-processing agreement, with the controller obligations resting on the customer. Where the GDPR's rules on automated decision-making apply to QA scoring of employees, the confidence flags, cited evidence, and human-review requirement above are the mechanisms that support meaningful human involvement.
• EU AI Act — customers deploying call analytics for EU workplaces should assess their obligations, including the Act's restrictions on emotion inference in the workplace. Relevant to that assessment: sentiment in Whizz Pulse is conversational sentiment derived from transcript text — what was said — not biometric or voice-tone emotion recognition. This is not legal advice; obligations depend on how and where you deploy.

6. What we do not do

• We do not make, answer, or record calls — the AI analyzes recordings customers upload, nothing else.
• We do not identify speakers biometrically and store no voiceprints; diarization separates voices within one recording only.
• We do not analyze voice tone, faces, or video, and we do not infer protected characteristics.
• We do not train models on call audio, transcripts, or analysis outputs.
• We do not score without evidence — every QA criterion score cites verbatim transcript quotes, and low-confidence results are flagged for human review.

7. Questions

If your compliance team needs specifics — the model used per stage (recorded on every analysis), data flows, subprocessors, or redaction behavior — write to support@whizztech.ai. Engineers answer this inbox.